ChatGPT Memory: What It Remembers About You and How to Control It

How ChatGPT Memory Works Now

ChatGPT Memory is a background process that reads across your conversations, files, and connected apps, decides which details are worth carrying forward, and writes them into a summary that gets attached to your future chats. You are no longer the one deciding what gets saved.

It is worth separating two things people routinely confuse. Chat history is the list of conversations in your sidebar. ChatGPT Memory is a separate store that sits above those conversations, holding conclusions drawn from them. They are managed in different places and they delete independently — which is where most of the trouble later in this guide comes from.

Saved memories versus the summary

Both systems still exist, and you can switch between them. The difference is who does the writing.

OpenAI’s own announcement of the rebuilt system, published on 4 June 2026 and titled “Dreaming”, describes the shift plainly: memory now updates itself as time passes, so “you’re going to Singapore in July” quietly becomes “you went to Singapore in July 2026” once the trip ends. Applied to work, that means an old project stops being described as current.

Which version are you on?

Open Settings, then Personalization, then Memory. If you see a paragraph-style summary with a “last updated” stamp at the top, you are on the new system. If you see a scrollable list of one-line entries instead, you are still on legacy saved memories.

The rebuilt system reached Plus and Pro users in the US first and has been rolling out to more countries and to Free and Go accounts since. If yours has not switched yet, everything in this guide still applies — you will just be deleting lines rather than correcting sentences.

What ChatGPT Memory Stores

ChatGPT Memory stores working context, not a transcript. Think of it as the notes a new colleague would take about you in their first month: what you do, who you write for, how you like things laid out, what you are currently working on. It is a paraphrase of you, written by a model, and it will not match your own wording.

• Role and industry — your title, your team, the audience you write for.
• Output preferences — brief length, list style, tone, words you never use.
• Company facts — headcount, working model, product lines, house terminology.
• Current work — the project you keep mentioning, the deadline you keep referencing.
• Tools — the software you name often enough for it to matter.

None of that is a problem while it stays accurate. The trouble starts when it does not, because ChatGPT applies old context with exactly the same confidence as new context.

Here is the part most guides miss, and it comes straight from the official OpenAI Memory FAQ: the summary does not show everything ChatGPT remembers. It is a high-level view of a broader synthesis, and some details are deliberately left off the page. If you want to know whether a specific fact is in there, the reliable move is to ask ChatGPT directly in a chat.

Memory also now draws on files you have uploaded and apps you have connected, which widens the surface considerably. If you have connected your email, calendar, or file storage to ChatGPT, those are inputs too.

The Seed–Audit–Prune Cycle

The Seed–Audit–Prune Cycle is a three-phase routine for keeping ChatGPT Memory accurate: seed the facts worth carrying, audit what actually landed, prune what has gone stale. It is the approach we teach at PromptPeakAI for any tool that holds context on your behalf, and it works because memory decays on a schedule — quarterly, roughly, for most professional roles.

The reason to run it as a cycle rather than a one-off setup is that ChatGPT Memory now updates itself. Anything you fix today can be quietly amended tomorrow by something you said in passing, so the audit pass is the part that actually keeps the system trustworthy over a year.

Do not wait for ChatGPT to guess what matters. Give it one deliberate briefing covering the four things that shape every piece of work you ask for: who you are, who reads your output, what your house rules are, and what you are working on now. The structure below follows the same Role, Goal, Context, Format logic we use for prompts.

Please remember the following as standing context for my work, and confirm what you have stored.

MY ROLE: [job title] at a [size] [industry] company. I mainly produce [3 main deliverables].

MY READERS: [who receives your work, e.g. "our executive team and external client contacts"].

MY RULES: Keep briefings under [X] words. Use [bullet points / short paragraphs]. Use [British / American] spelling. Never use the words [list 3-5 words you hate].

CURRENT WORK: [project name and what it is], running until [rough date].

Do not store anything from this message that is not listed above.

Now check the result, because what ChatGPT stored will not match what you typed word for word — it paraphrases. Open Settings, Personalization, Memory and read the summary top to bottom. Then ask directly, since the summary is only a partial view.

List everything you currently remember about my professional work — my role, my employer, my responsibilities, my writing preferences, my current projects, and any clients or colleagues you have stored.

Group them under those headings. For each item, tell me roughly when you think you learned it.

Flag anything you are holding that looks out of date or that you are unsure about.

On the new system you do not delete lines — you correct the summary. Type the correction into the box at the bottom of the memory summary, or highlight the specific sentence that is wrong and fix it in place. On legacy saved memories, you delete the offending entry outright.

Update what you remember about my work. These details have changed:

NO LONGER TRUE: [old job title, old employer, finished project, retired policy].

NOW TRUE: [current equivalent for each].

STOP REFERENCING: [anything you want dropped entirely, e.g. a former client].

Confirm each change, and tell me if any of these still appear somewhere you cannot update from here.

Where Each Piece of Context Belongs

Memory is one of four places context can live, and most of the frustration people have with ChatGPT comes from putting things in the wrong one. Permanent rules do not belong in memory; evolving facts do not belong in Custom Instructions.

The practical split: your job title, your reading audience, and your banned-word list are stable for years, so they belong in ChatGPT Custom Instructions where you control the text directly. The client you are currently onboarding belongs in memory, where it can update itself and then quietly age out.

If your worry is contamination — client work bleeding into personal chats, or one account’s context showing up in another’s deliverable — the actual fix is project-only memory. When you create a new ChatGPT Project, you can scope its memory to that project alone. Two warnings: you can only choose it when the project is created, and once set it cannot be reverted.

The Chat Deletion Trap

Deleting a chat does not delete the memories created during that chat. This is the single most expensive misunderstanding about ChatGPT Memory, and it catches careful people constantly — usually the ones who deleted the conversation precisely because they knew they had shared too much.

The reason is structural. ChatGPT Memory lives at the account level, not inside the conversation that produced it, so removing the conversation removes the source but leaves the conclusion behind.

Keeping Work Data Out of Memory

The safest rule is the simplest one: if a detail would need a confidentiality agreement to leave the building, it should not be sitting in a consumer AI account’s persistent memory. Sort your inputs before you paste, not after — ChatGPT Memory has no concept of which facts are yours to share.

Separately from memory, check whether your conversations are being used to train models. On personal plans that is a toggle: Settings, Data Controls, “Improve the model for everyone”. On ChatGPT Business and Enterprise workspaces, OpenAI excludes workspace conversations and memories from model training by default — which, as most IT leads will tell you, is the actual reason to put a team on a business plan rather than reimbursing individual subscriptions. Our guide to using ChatGPT safely at work covers the wider policy question.

For genuinely sensitive one-off tasks, use Temporary Chat. It does not read your existing memories and does not create new ones, and those conversations are not used for training.

[Start a Temporary Chat first, then paste this.]

You are helping me summarise sensitive internal notes. Work only from the text I paste below.

Rules for this session:
- Refer to people by initials only. I have already removed full names.
- Do not infer or fill in any detail I have not given you.
- Do not save anything from this session.

TASK: Summarise the notes below into [3 coaching points / 5 decisions and owners].

NOTES:
[paste your anonymised notes]

When ChatGPT Memory Goes Wrong

Most ChatGPT Memory problems look like the model being stubborn or forgetful. They are almost always a mismatch between what you think it stored and what it actually stored. The book icon underneath any response opens a sources panel showing which memory shaped that answer — start there before you start deleting anything.

That last row deserves emphasis. A wrong memory does not produce a hedged answer — it produces a confident one, which is why it slips past review. If you have read our explanation of why ChatGPT states wrong things confidently, stale memory is the same failure with a traceable cause and a two-minute fix.

Forget everything you currently remember about [topic, client, or project name].

Do not replace it with an assumption. Treat it as something you have no information about until I tell you otherwise.

Then tell me plainly: is there anything about [topic] you are still holding that you cannot remove from here?

What ChatGPT Memory Still Can’t Do

ChatGPT Memory is useful and genuinely limited, and being clear about the limits is what keeps it from quietly embarrassing you in front of someone who matters.

None of that argues against using it. It argues for a maintenance habit — the same one you would apply to any system that stores context on your behalf, which is increasingly all of them. If you want the wider version of this thinking, context engineering is the discipline this fits inside, and the same audit logic applies to Claude’s memory feature if you work across both tools.

Run the seed pass this week. Put a quarterly reminder in your calendar for the audit and prune. Then get on with the work, and let the tool carry the context you deliberately gave it — which is, after all, the point. Setting this up alongside a proper ChatGPT work configuration takes an afternoon and pays back every week after.